Conditional Generative Flow-Induced Variational Autoencoder for Plasma Instances Augmentation

In this article, we propose a novel conditional generative flow-induced variational autoencoder (CGlow-VAE) model to address the critical challenge of the small sample issue in plasma instances. This approach integrates variational inference with conditional generative flow, establishing a bidirectional mapping between high-dimensional data and their corresponding labels. Specifically, the encoder network maps the input data to a structured latent distribution, while the conditional generative flow module systematically optimizes the log-likelihood of the observed labels through a series of invertible transformations, treating them as conditional variables. This process effectively captures the complex nonlinear coupling between plasma characteristics and measurement outputs. Based on this framework, the decoder reconstructs input data, ensuring that the generated data maintains distributional consistency with the original data. The trained conditional flow is then used to reverse-generate the corresponding label data. To evaluate the effectiveness of our proposed method, we conducted a comprehensive experimental assessment using the plasma flash imaging dataset and the optical emission spectroscopy (OES) dataset. In addition, we compared our approach against existing state-of-the-art methods to ensure a thorough performance evaluation. The results demonstrate that the proposed CGlow-VAE achieves significant improvements in sample generation while effectively mitigating the issue of small samples through data augmentation, thus enhancing the generalizability of the model.